DE4215179A1 - Synaptic processor for neural network - initialises input expectation limited by maximum and neutral expectations then modifies expectation using increase and reduction terms - Google Patents

Abstract

The synaptic processor for generating a result signal as a result of an input signal contains an arrangement for initialising an input expectation which is limited by maximum and neutral expectations. It contains an arrangement for receiving the input signal and modifying the input expectation. The net modification of the input expectation contains an increase and a reductio term. The increase term is a function of the input signal and a first increasing gradient constant. The reduction term is a function of a first decreasing gradient constant independent of the increaseing constant. A further arrangement generates the result signal from the input expectation. USE - For use as component of neutral network.

Description

The invention relates generally to nerve or neural networks and in particular on a processor to build artificial neurons. Interest in the information Brain's manufacturing skills have existed since the late 19th century. However, only recently have artificial networks based on neurons acquired the skills to to be really useful. For example, find artificial Neural networks uses in vision and speech recognition systems, in signal processing and robot application.

The research and development of neural networks has grown up Try to develop the processes and functions of the brain replicate. Therefore, much of the terminology has biological Origins.

The basic component of a neural network is a neuron. A Neuron can be viewed as a weighted adder. A Neuron has a number of inputs. Every input variable will multiplied by a weighting value and the products are added. The weightings can be positive (stimulating) or be negative (inhibitory). The output of the neuron is a radio tion of the sum of the products.

A (single) neural network is known as that Single layer or single level network. Single-layer networks are useful for solving linear separable problems. An application of a single layer network factory is the pattern recognition.

In the simplest case, each neuron is designed for identification a specific pattern in the input signals.  

Every neuron in the network has the same inputs. The Each neuron's output is either a "hit" or a "miss". The sum of the products for each Neuron is compared to a threshold. If the sum of a neuron is greater than the threshold, then it has this pattern is recognized and signals a "hit". The Number of expenses of the network is equal to the number of Patterns to be recognized, and therefore the number of Neurons. Only one neuron gets a "hit" at the same time signal (if its pattern appears in the entries).

In a more complex case the output of the neuron is not one "Hit" or "missed". Rather, it is a graded one Scale value that shows how close the input pattern is to the Is neuron pattern.

For more difficult problems there is a multi-layer network needed.

The weights of a neuron can either be fixed or customizable be. Customizable weights make a neural network more flexible. Customizable weights are made through learning laws or rules modified. Most learning laws are based on association activity. Based on a learning rule developed by Donald O. Hebb Was proposed in 1949 (Hebb’s rule), has learning theory generally believed that the essence of learning phenomena is a Association between two or more signals. In Hebb's rule, for example, is the asso with one entry graced weighting increases when both the input line and the output line is active at the same time. There were many Va rations on this topic, but one way or the other they are guided most of the neural learning rules from this basis.

More recently, other learning laws have become associative Rules busy on simultaneous input signals based. With this approach, the weighting becomes a Input increases if this input and a certain after  cash receipt both within a limited time window are active.

It is now generally accepted that, for example, in the active pattern recognition and classification the correlation between multiple input signals or input signals and Output signals is the adaptive mapping mechanism forms. Any system that would "learn" a pattern to it know and / or classify a category, it must Encrypt or encode correlations. Hence the most, if not all, work on one on association based mechanism.

However, each of these learning rules has proven to be limiting posed. Each rule works well for a given spec fish type or group of problems, but none of them have hitherto considered sufficient general or adaptable exposed to for the wide range of problems to be applied in this area.

The word adaptive is the key. Adaptation or adaptives Adaptive response AR is the process of change the response of the network to meet the requirements of to meet new conditions.

Such changes do not occur in a biological system carried out on the first experience of a change. Rather, the change must be frequent and / or over a long period Time frames occur before the organism becomes one undergoes significant change. Biological systems tend about being very conservative in this regard.

The development of muscle tissue in an athlete or sport It provides a good example of this aspect of adaptive Speak to. Let's consider a wannabe weight lifter who just starting this sport.  

After a week of practice, he may be able to do considerably more Lifting weight than he could at the beginning. Note that the Type of exercise is periodic with long intervals (in Ver at the time of the exercise itself) between episodes. In short in time, the body keeps the memory of the episodes physical requirement that was placed on the muscles.

This memory increases with the continuation of the exercise. If the weight lifter continues for a week, he will notice that lifting the same weight a bit at the end of this week will be easier than at the beginning.

If he continues training for a month, he will find that he can lift even more weight with the same apparent weight the effort that was needed for less weight in the Early days.

Continued regular training turns into physical changes changes in the muscles. Muscle packs swell when new tissue is built up around the to help meet continued requirements. A kind of Long-term memory will emerge, making this new Tissue will remain in place for some time, even if that Training has been abandoned for a while.

Why doesn't the body produce this new tissue at first or even second requirement episode? The reason is that biological tissues have evolved to to meet the stochastic nature of the environment. They are inherent (inherently) conservative and only hold the amount of an speak upright, as is "normally" necessary, to master everyday life. Only after the body like who was repeatedly exposed to the new level of requirements speaks he starts by creating new muscle fiber.

What happens to the athlete who loses training? To a certain extent, it depends on how long he is has trained. If he stops after a week, then he will  about his ability within another week returned to training. If, however, after 6 months Training stops, it can take many months for the new one Capacity wilts back to the level before training. The athlete should complete a training program within several months after resuming it is indeed true apparently that he is peak at a relative level short time compared to the initial process. This is due to the long-term memory effect of the adaptive response.

Indeed shows the adaptive response of biological Systems that learning is not strictly a correlation is between two signals. Rather, it is a more general one Consideration that learning on encrypting Information based in the changing over time Signal inherent in a single input channel.

The present invention is directed to a processor to be provided as a building block of a neural network based on based on this rather basic learning law.

According to one aspect of the present invention, a synap tical processor provided for generating a result signal due to an input signal. The synaptic processor includes first and second expectations. The synaptic The processor modifies the first expectation as a function of the Input signal and the second expectation and modified the second expectation as a function of the first expectation. The Result signal is determined based on the first expectation.

According to a further aspect of the present invention, a Processing element provided. The processing element includes a synaptic input processor to generate a Result signal due to an input signal. The synaptic Input processor includes first and second expectations. The synaptic processor modifies the first expectation as one Function of the input signal and the second expectation and  modifies the second expectation as a function of the first Expectation. The result signal is determined as a result of the first Expectation. The processing element generates an output signal due to the result signal and a threshold signal.

Fig. 1 is a block diagram of a synaptic processor or Adaptrode;

Fig. 2 is a schematic representation of the input modification means of the synaptic processor in Fig. 1;

Fig. 3 is a schematic representation of the result signal generating means of the synaptic processor of Fig. 1;

Fig. 4 is a graphical representation showing the response of the input expectation of a single-level adapter with linear gain to an input signal with a high input expectation;

Fig. 5 is a graph showing the response of the input expectation of a single level and linear gain gain to an input signal of an adapter with a low input expectation;

Fig. 6 is a graph showing the response of the input expectation of a single level, non-linear gain adaptrode to a high intensity, short duration input signal;

Fig. 7 is a graph showing the response of the input expectation of a single level, non-linear gain adapter to an input signal of high intensity and medium duration;

Fig. 8 is a graph showing the response of the input expectation of a single level, non-linear gain adaptrode to a high intensity, long duration input signal;

Fig. 9 is a graph showing the response of input expectation of a single level, non-linear gain adaptrode to a saturating input signal;

Fig. 10 is a graph showing the response of the input expectation of a single level, non-linear gain adaptrode to a stochastic input signal;

Fig. 11 is a graph showing the response of the input expectation of a single level, non-linear gain adaptrode to a sequential input signal;

Fig. 12 is a graph showing the response of the input expectation of a single level linear gain adaptrode to a high intensity, short duration input signal;

Fig. 13 is a graph showing the response of the input expectation of a single-level adapter with linear gain to an input signal of high intensity and of medium duration;

Fig. 14 is a graph showing the response of the input anticipation of a Adaptrode with a single plane with linear amplification to an input signal with high intensity and of long duration,

Fig. 15 is a graphical representation showing the response of the input expectation of a single level linear gain adapter to a saturating input signal;

Fig. 16 is a graph showing the response of the input expectation of a single level linear gain adapter to a stochastic input signal;

Fig. 17 is a graph showing the response of the input expectation of a single level linear gain adapter to a sequential input signal;

Figure 18 is a block diagram of a multi-level synaptic processor or adapter;

Fig. 19 is a graphical representation of the multi-level synaptic processor of Fig. 18;

Fig. 20 is a graphical representation of a multilevel synaptic processor using a pass or gate function;

Figure 21 is a schematic representation of the expectation modifying means of a three level synaptic processor;

Fig. 22 is a schematic illustration of the efficacious signal generating means of Fig. 18;

Fig. 23 is a schematic illustration of the hurdle sign source of Fig. 22;

Fig. 24 is a graph showing the response of the expectations of a three-level, non-linear gain adapter to an input signal of high intensity and short duration;

Fig. 25 is a graph showing the response of the expectations of a three-level, non-linear gain adapter to an input signal of high intensity and medium duration;

Fig. 26 is a graphical representation showing the response of expectations of a three-level, non-linear gain adapter to an input signal of high intensity and long duration;

Fig. 27 is a graph showing the response of the expectations of a three-level, non-linear gain adapter to a high intensity saturating input signal;

Fig. 28 is a graphical representation showing the response of the expectations of a three-level adaptrode with non-linear gains to a stochastic input signal;

Fig. 29 is a graphical representation showing the response of the expectations of a three-level, non-linear gain adapter to a sequential input signal;

Fig. 30 is a graphical representation showing the response of expectations of a three-level, non-linear gain adapter to an input signal of high intensity and short duration, with the second expectation provided with a gate;

Fig. 31 is a graphical representation showing the response of expectations of a three level, nonlinear gain adapter to a high intensity, medium duration input signal with the second expectation gated;

Fig. 32 is a graph showing the response of expectations of a three-level, non-linear gain adapter to an input signal of high intensity and long duration with the second expectation gated;

Fig. 33 is a graphical representation showing the response of the expectations of a three-level, non-linear gain adapter to a high intensity saturating input signal with the second expectation gated;

Fig. 34 is a graph showing the response of the expectations of a three-level, non-linear gain adapter to a stochastic input signal, the second expectation being gated;

Fig. 35 is a graphical representation showing the response of the expectations of a three-level, non-linear gain adapter to a sequential input signal with the second expectation gated;

Fig. 36 is a graph showing the response of expectations of a three-level adapter to a high intensity, short duration input signal, the first expectation having a linear gain;

Fig. 37 is a graph showing the response of the expectations of a three level adapter to a high intensity, medium duration input signal, the first expectation having a linear gain;

Fig. 38 is a graph showing the response of input expectations of a three-level adapter to a high intensity, long duration input signal, the first expectation having a linear gain;

Fig. 39 is a graph showing the response of expectations of a three level adapter to a high intensity saturating input signal, the first expectation having a linear gain;

Fig. 40 is a graph showing the response of the expectations of a Adaptrode with three levels on a stochastic input signal, wherein the first expectation has a linear gain;

Fig. 41 is a graph showing the response of the expectations of a three level adapter to a sequential input signal, the first expectation having a linear gain;

Fig. 42 is a graphical representation showing the response of the expectations of a three-level adapter to a high intensity, short duration input signal, the first expectation having a linear gain and the second expectation having a gate;

Fig. 43 is a graph showing the response of the expectations of a three-level adapter to a high intensity, medium duration input signal, the first expectation having a linear gain and the second expectation having a gate;

Fig. 44 is a graph showing the response of the expectations of a three-level adapter to a high intensity, long duration input signal, the first expectation having a linear gain and the second expectation having a gate;

Fig. 45 is a graph showing the response of the expectations of a three-level adapter to a high intensity saturating input signal, the first expectation having a linear gain and the second expectation being gated;

Fig. 46 is a graphical representation showing the response of the expectations of a three-level adapter to a stochastic input signal, the first expectation having a linear gain and the second expectation being gated;

Fig. 47 is a graphical representation showing the response of the expectations of a three-level adapter to a sequential input signal, the first expectation being a linear gain and the second expectation being gated;

Fig. 48 is a block diagram of a processing element with an input adapter arrangement, an add and output processor and a threshold processor;

Fig. 49 is a graphical representation of the add and output processor of Fig. 48;

Fig. 50 is a schematic illustration of the multiplication grid of Fig. 49;

Fig. 51 is a schematic illustration of the threshold processor of Fig. 48;

Figure 52 is a graphical representation of two synaptic processors, with the output of one used as a gate for the other;

Fig. 53 is a legend for Figs. 58-64;

Figure 54 is a block diagram of an associative network with two neuristors showing the association in conditioned response;

Fig. 55 is a block diagram of the network of Fig. 54 in which a conditional stimulus has been associated with an unconditional stimulus / response;

Fig. 56 is a block diagram of the network of Fig. 55 after a period without stimulus;

Fig. 57 is a block diagram of the network showing the conditional stimulus / response association of Fig. 56;

Figure 58 is a block diagram of the network of Figure 57 showing the effect of an opposite unconditional stimulus;

Figure 59 is a block diagram of the network of Figure 58 showing the effect of a subsequent intermittent stimulus / response association; and

FIG. 60 is a block diagram of the network of FIG. 63, showing the long-term association between the conditional response and the first non-conditional stimulus.

Best mode for carrying out the invention

Referring to Fig. 1 is shown an embodiment of a synaptic processor 102 or a Adaptrode with a single plane for use in a neural network.

It should be noted that part of this application is by Contains copyrighted material.

Traditionally, neural networks have been run under Use a wide range of technologies. This tech Technologies can be analog or digital. The actual order Settlement changes with the application, based on example wise on network complexity. During analog circuits provide an inexpensive but quick application larger networks arduous and some accuracy goes lost. More complex networks are generally based on All purpose computers or specially designed hardware. For this reason, it is practical to use the present invention to describe mathematically. However, it should be noted that the present invention does not apply to such a mathematical  Execution or even analogue or digital execution is limited.

The synaptic processor 102 receives an input signal X and generates a result signal r. The result signal is a function of an input expectation W ⁰ , as described below.

In parts of the following discussion there is a distinction made between a continuous system and a dis cretan system. In a continuous system, the System parameters continuously updated and are called continuously variable. In one discrete system, the system parameters are in discrete Time intervals, typically referred to as Δt, on the brought up to date.

Furthermore, the expression intensity such as high intensity or medium intensity is used in the description of the input signal. For example, in a discrete system in which the input signal is formed as a series of pulses with a set width, a high intensity input signal consists mainly of pulses. The single-level adapter 102 comprises means 104 for initial setting (initialization) of the input expectation. The input expectation W ⁰ is limited by a maximum expectation W ^M and an equilibrium expectation W ^E. In one embodiment, the input expectation W ^{0 is given} the initial value equal to W ^E , typically zero (0).

Means 106 receives the input signal and determines one Size of the input signal. The size of the input signal is depending on the exact nature of the signal. For example, the input signal can be in a continuous System have a changing frequency. The input signal can also be a changing continuous voltage or Have electricity. If the neural network is digital, the input signal is a logic level low (0) or on logic level high (non-zero).  

Means or a process unit 108 modifies the input waiting W ⁰ . The means 108 calculate a net modification for the input expectation. The net modification is added to the previous input expectation to form a new input expectation. The net modification has an enlargement term and a reduction term. The enlargement term tends to increase the input expectation and the decrease term tends to decrease the input expectation. The net effect of the two terms depends on their relative size.

The input expectation W ⁰ represents the current expectation of the adapter with regard to the intensity of the input signal. The adapter adapts or learns from the input signal. In other words, the Adaptrode keeps a memory of the waiting for his past. His expectation of the future input signal is based on this past.

The magnification term is a function of the input signal and tends to "pull up" the input expectation. The sales The term is not a function of the input signal and provides the adaptrode with a built-in preset, to expect them to send a weak signal in the long run Will be received.

In one embodiment, the increase term is a function of the input signal and the decrease term is independent of the input signal. In another exemplary embodiment, the increase term is a function of a rise constant α ⁰ and the decrease term is a function of a decrease constant δ ⁰ . Furthermore, the rise and fall constants are independent. In another embodiment, the magnification term is a function of only one input signal.

In a first embodiment, the input expectation modified as follows:

W⁰ = W⁰ _p + X · α⁰ · [W ^M - W⁰ _p ] - δ⁰ · [W⁰ _p - W ^E ] (Eq. 1)

in which

W⁰ = the new input expectation,
W⁰ _p = the previous input expectation,
X = the input signal,
α⁰ = the climb constant,
W ^M = the maximum expectation,
δ⁰ = a drop constant, and
W ^E = equilibrium expectation

is.

In the first embodiment, the enlargement term is one non-linear function. It is therefore said that an adapter with a level corresponding to the first embodiment has non-linear gain.

In a second embodiment, the input expectation modified as follows:

W⁰ = W⁰ _p + X · α⁰ - δ⁰ · [W⁰ _p - W ^E ] (Eq. 2)

In the second embodiment, the magnification term is a linear function. Therefore, it is said that the adapter according to the second embodiment has a linear gain. Furthermore, the input expectation W ^{0 is} not expressly limited to a maximum value. However, simulation has shown that Equation 2 achieves a maximum steady state equilibrium value when a continuous input signal is present for an appropriate period of time. The exact maximum value is a function of the equation parameters, the intensity of the input signal and the size of the input signal.

In both linear and non-linear gain models, the rate constants are set so that the rate of fall is longer than the rate of climb. That is, α ⁰ is larger than δ ⁰ . Referring to FIG. 2, the input erwartungsmodifizierenden means comprise means 202 for 108 men Bestim the magnification term and means 204 for determining the reduction term.

From equation 1 the magnification term is:

X · α⁰ · [W ^M - W⁰ _p ].

In the preferred embodiment, the magnification term is positive and is added to the input expectation. The means 202 determining the enlargement term comprise a delay block 206 for generating a delayed or previous input expectation W ⁰ _p . The delayed input expectation (W ⁰ _p ) is subtracted from the maximum expectation WM by a first adder (summer) 208. A first multiplier 210 multiplies the output of the first adder 208 by the rising constant α ⁰ .

Similarly, the reduction term in Equation 1 is defined as:

δ⁰ · [W⁰ _p - W ^E ].

In the preferred embodiment, the decrease term is positive and is subtracted from the input expectation. The means 204 determining the reduction term comprise a second adder 212 for subtracting the equilibrium expectation from the delayed input expectation. A second multiplier 214 multiplies the output of the second adder 212 by the decay constant δ ⁰ . A third multiplier 216 multiplies the output of the first multiplier 210 by the input signal X. A third adder 218 adds the output of the third multiplier 216 and subtracts the output of the second multiplier 214 from the delayed input expectation to form the new input expectation W ⁰ .

Referring back to FIG. 1, means 110 generates a result signal r as a function of input expectation W ⁰ . In one embodiment, the result signal is a linear function of the input expectation.

In a further exemplary embodiment, the result signal is a non-linear function of input expectation. For example is in a continuous system:

r = ^KW⁰ e ^{- λτ} ( ^Eq.3A )

in which

r = the result signal,
k = a first constant,
W⁰ = the input expectation,
λ = a second constant, and
τ = a predetermined time differential

is.

The time differential τ is determined as the running time minus the time at which the size of the input signal is non-zero has become. Another example is for a discrete System:

r (t) = kW⁰,

if X <0, or otherwise

r (t) = r (t-Δt) - λ · r (t-Δt) (Eq. 3B)

in which

r = the result signal,
k = a first constant,
W⁰ = the input expectation, and
λ = a second constant

is.

With reference to FIG. 3, the resultant signal-generating means 110 are shown for converting the equation 3B. The means 110 comprise a fourth multiplier 302 for multiplying the input expectation W ⁰ by a constant k. A second delay block 303 passes the result signal to a fifth multiplier 304 and a fourth adder 306. The fifth multiplier 304 multiplies the delayed result signal by a constant. The fourth adder 306 subtracts the output of the fifth multiplier 304 from the delayed result signal. Decision block 308 compares the input signal X to a threshold (e.g. zero). If the input signal is greater than the threshold, the output of fourth multiplier 302 is passed through control block 310. Otherwise the output of the fourth adder 306 is passed on.

A graphic representation of the adapter 102 is used to illustrate certain operating states of the adapter 102. Referring to Figs. 4 and 5 is applied an input signal X to the input expectation modifying means 108. The enlargement and reduction dates are determined and used to determine the new input expectation. The result signal generating means 110 (response unit) determine the result signal as a function of the input expectation.

The adapter electrode 102 of FIG. 4 has a high input expectation, as is illustrated by the bar 402. In contrast, the bar 502 of the adapter 102 of FIG. 5 has a relatively low input expectation. Similar representations are used throughout the following discussion to illustrate important features of the present invention.

With reference to FIGS. 6 to 11 the input expectation W ⁰ a Adaptrode 102 having a single plane and non-linear gain playback signal waveforms shown for a number of different A. In each example, the rise constant is α ⁰ and the decay constant is δ ⁰ 0.25 and 0.125, respectively. It should be noted that these constants have been chosen for purposes of illustration only. The maximum expectation W ^M is set to one and the equilibrium expectation W ^E is set to zero. The input expectation starts from zero. Furthermore, the input signal has a size of 1.0.

The graphs used in the following discussion used to respond to input expectations modifying means 108 for a discrete system  that is, the input expectation is at discrete time intervals modified from δt. For a single adapter adapter 102 Level and non-linear gain becomes the input expectation determined by:

W⁰ (t) = W⁰ (t-Δt) + X (t) · α⁰ · [W ^M -W⁰ (t-Δt)] - δ⁰ · [W⁰ (t-Δt) - W ^E ]. (Eq.4A)

With reference to FIGS. 6-17 and 24-47 the input signal is a series of pulses of constant width and size. The input signal is shown as a bar at each display below. The bar indicates that the input signal is not zero. It should be noted that while the representations in Figures 6-17 and 24-47 are based on discrete data, the data is represented by a line graph. The line diagram is particularly useful when examining the responsiveness of the adapter.

In Fig. 6 = 0 at time t, the input signal is equal to zero and the input expectation is also zero. At time t = 5, the input signal is non-zero and has a size of 1.0. The input remains non-zero until t = 9.

As shown, while the input is non-zero, the increase term is greater than the decrease term and the input expectation increases exponentially. After the input signal becomes zero again, the input expectation drops to the equilibrium expectation W ^E.

In Fig. 7, an input signal of medium duration is applied to the adapter. As shown, an input signal of magnitude 1.0 is applied at time t = 5 and becomes zero again at t = 14. It should be noted that the net modification actually decreases, although the input expectation increases over the period during which the input signal is not zero. This is attributed to the fact that the magnification term is a function of the difference between the maximum expectation and the current expectation and because of the increasing effects of the diminishing term.

This effect is further shown in Fig. 8, in which an input signal of high intensity and long duration is applied to the adapter. The input expectation seems to be approaching a certain value.

Indeed, as shown in Figure 9, the input expectation reaches a saturation point or steady state or equilibrium value with an input signal of maximum intensity and constant size. In this situation, the enlargement term and the reduction term balance each other and the net modification is zero.

In Fig. 10, as shown, a stochastic or random input signal has been applied to the adapter 102. If the input signal is non-zero, the increase term will make the net modification positive, and if the input signal is zero, the decrease term will make the net modification negative.

In Fig. 11, a sequential input signal has been applied to the adapter. That is, the input signal is non-zero for a first short period or period (5 to 10), is zero for a second period (11 to 24) and then is not zero for a third period (25 to 29). The resulting input expectation shows the learning ability of the single-level adapter. During the first and second periods, the input expectation rises and falls as expected and as described above. At t = 25 the input signal becomes non-zero and the input expectation begins to rise again. However, it increases based on its current or current value.

The rise and fall constants will affect the type of enlargement and reduction rates of the input expectation. For example, if the rise and fall constants are of a similar size, the increase and decrease rates will be similar to the input expectation. If, on the other hand, the decay constant is relatively smaller than the rise constant (δ ⁰ «α ⁰ ), the adapter will learn" quickly "and slowly" forget ".

With reference to FIGS. 12-17, the input expectation W ⁰ a Adaptrode 102 having a single plane and having a linear Ver strengthening shown for a number of different input signals. In each example, the climb constant α ⁰ and the decay constant δ ^{0 are} 4.0 and 0.125, respectively. The equilibrium expectation W ^E is set to zero and the input expectation starts at zero.

Figures 12-17 illustrate the response of the input expectation modifiers for a discrete system. Therefore, the input expectation for a single level, linear gain adapter is determined by:

W⁰ (t) = W⁰ (t-Δt) + X (t) · α⁰ - δ⁰ · [W⁰ (t-Δt) - W ^E ]. (Eq.4B)

In Fig. 12 an input signal of high intensity and short duration is applied to the Einzelebenenadaptrode 102nd In Figs. 13 and 14, an input signal of high intensity and of medium or long duration is applied to the Adaptrode 102nd It should be noted, however, that although the enlargement term is defined as a linear function, the input expectation due to the reduction term is not a linear function.

In FIG. 15, an input signal of maximum intensity and with a constant size (1.0) is applied to the single-level adapter 102. As with the single-level, non-linear gain adaptrode 102, the input expectation of the single-level, linear gain adaptrode reaches an equilibrium value. Note, however, that the input expectation W ⁰ by Equation 4 is not limited to a maximum value. Rather, the equilibrium value is a function of the adapter parameters, the intensity of the input signal and the size of the input signal. For a given adapter, the equilibrium value would reach a maximum expectation if an input signal of maximum intensity and maximum size was applied for a suitable duration.

In Fig. 16, a stochastic input signal to the A 102 is applied zelebenenadaptrode. In Fig. 17, a sequential input signal is applied to the single-level adapter. Both diagrams show that the single-level adaptrode with a linear gain has a similar ability to learn as the adaptrode with non-linear gain.

These examples show the "learning" (climb -) and "forgetting" (Decay -) aspects of the single level adapter. Furthermore, because the rise and fall constants are independent, the "learning" - and "forgetting" aspects of each adapter set independently will. That means both can be fast, both can be slow, or any combination as desired Properties.

The single-level adapter shows memories skills in a time domain or area. By allowing of changes in equilibrium expectations (for the Input expectation) there is the one shown by the adapter Memory ability in multiple time domains. Such a Adaptrode becomes multi-level or Multi-level adapter called.

Referring to Figure 18, a synaptic processor 1002 or multi-level adapter has a variety of expectations. The expectations are generally denoted by W ^l , where 0 l is L.

A mean 1004 gives the expectations W ^l initial values. The expectations are limited by a maximum expectation W ^M and an equilibrium expectation W ^E. In general, the expectations have the relationships:

W ^M W⁰ W¹ W ^L W ^E.

A means or process unit 1006 (1) modifies the expectation W ¹ . The modifying means 1006 (l) determine a net modification and add the net modification to the expectation W ^l . The net modification has an enlargement term and a reduction term. Generally, the magnification term is a function of the previous expectation W ^l-1 if l 0. If l = 0, the magnification term is a function of the maximum expectation W ^M. The reduction term is a function of the following expectation W ^{l + 1} if l <L or W ^E if l = L.

In a first embodiment, the first expectation W ^{0 is} a function of the size of the input signal.

In a second embodiment, the first expectation W ^{0 has} a linear gain (as described above).

In a third embodiment, the first expectation W ^{0 has} a non-linear gain (as described above).

In a fourth exemplary embodiment, the increase term of each expectation W ^{l is} a function of a rise constant α ^l and the decrease term is a function of a fall constant δ ^l .

In a fifth exemplary embodiment, the enlargement term of each further expectation W ^l (l <0) is a function of the difference between the previous expectation W ^l-1 and the current expectation W ^l .

In a sixth embodiment, the reduction term of each expectation W ^{l is} a function of the difference between the current expectation W ^l and the following expectation W ^{l + 1} or a function of the difference between the current expectation W ^l and the equilibrium expectation W ^E when l = L.

In a seventh embodiment, the first expectation modified according to:

W⁰ = W⁰ + X · α⁰ · [W ^M -W⁰ _p ] - δ⁰ · [W⁰ _p - W¹] (Eq. 5A)

where W ⁰ _{p is} the previous first expectation.

In an eighth embodiment, the first expectation modified according to:

W⁰ = W⁰ _p + X · α⁰ - δ⁰ · [W⁰ _p - W¹] (Eq. 5B).

Every further expectation (W ^l for l <0) is modified according to:

W ^l = W ^l _p + α ^l · [W ^l-1 - W ^l _p ] - δ ^l · [W⁰ _p -W ^{l + 1} ] (Eq. 6)

where W ^l _{p is} the previous expectation, W ^{l + 1 is} the next level expectation if l <L, or the equilibrium expectation W ^E if l = L, and where W ^{l-1 is} the previous expectation .

As in the single level adapter, the rate constants for each level are typically set so that the decay rate is longer than the rise rate. That means α ^l <δ ^l . Furthermore, in a multi-level adapter, each of the rate constants for each level (l) is set so that the rise and decay rates of one level (l) go over a longer time frame than the rise and decay rates of the previous level (l-1). This means:

α⁰ α¹ α ^l α ^L and d⁰ δ¹ δ ^l δ ^L.

Means 1008 generate the result signal as a function of the first expectation W ⁰ . As described above, the result signal can be a linear or non-linear function of the first expectation W ⁰ .

Referring to FIG. 19, the multi-level adapter 1002 is graphically shown as shown. The graphic representation shows several important features of the multi-level adapter 1002. First, the first expectation W ^{0 is} the only expectation that is directly influenced by the input signal X. Second, the result signal r is only a function of the first expectation W ⁰ .

Also, the size of the bars 1102 (l) is a sign of the right relative size of the corresponding expectation. Furthermore, the Arrows indicate that every expectation is both positive (by Ver enlargement term) as well as negative (due to the reduction term) being affected.

With reference again to FIG. 18, in a further exemplary embodiment, the multi-level adapter 1002 comprises means 1010 for receiving a second input signal H ¹ and for generating an enable or potentiation signal P ¹ accordingly. The enable signal is associated with one of the expectations for l <0.

In a first exemplary embodiment, the initiating means 1004 give an initial value to a pass or gate signal g ^l . The means 1010 generating the enable signal compare the second input signal H ¹ with the pass signal and in response generate the enable signal P ¹ . For example, in a discrete system P ^l can be determined by:

p ^l = 1 if H ^l <g ^l , or otherwise p ^l = 0. (Eq. 7)

In a continuous system, P ^{l is} a function of the difference between the second input signal H ^l and the pass signal g ^l .

In a second exemplary embodiment, the second input signal H ^{l is} determined as a function of the sum of a large number of interaction signals h ⁿ , where N denotes the number of interaction signals and where 0 n is N. The interaction signals are described below.

In a third embodiment, the pass signal g ^{l is} changed as a function of a parameter of the adapter 1002. For example, the pass signal can be a function of W ⁰ or of the result signal r.

With reference to Fig. 20 has in the preferred Ausführungsbei play any expectation (except for l = 0) associated through laßssignal g ^l. As shown, the enable achieved by the enable signal P ¹ only affects the magnification term or the "learning" of the expectation. The reduction term or the "forgetting" is not affected.

Therefore, equation 6 for l <0:

W ^l = W ^l _p + p ^l · α ^l · [W ^l-1 - W ^l _p ] - δ ^l · [W⁰ _p - W ^{l + 1} ] (Eq. 8)

where W ^{l + 1 is} the next expectation if l <L, and the equilibrium expectation W ^E if l = L.

If P = 1 in a discrete system, it is said that the Er Possibility signal is in a "permission" state since learning is permitted. If P = 0, it is said that the enabling signal is in a "prevention" state since learning is not is permitted.

With reference to FIG. 21, the expectation modifying means 1006 (0), 1006 (1), 1006 (2) for a multi-level adapter with three expectations (L = 2) are shown in further detail according to an embodiment of the present invention.

The first expectation modifying means 1006 (0) comprises means 1302 (0) for determining the increase term and means 1304 (0) for determining the decrease term of the first expectation. The enlargement term determining means 1302 (0) comprise a third delay block 1306, which generates a delayed or previous first expectation W ⁰ _p . The delayed first expectation (W ⁰ _p ) is subtracted from the maximum expectation W ^M by a fifth adder 1308. A sixth multiplier 1310 multiplies the output of the fifth summer 1308 by a first rise constant α ⁰ .

The enlargement term determining means 1304 (0) comprise a sixth adder 1312 for subtracting the second expectation W ¹ from the delayed first expectation W ⁰ _p . A seventh multiplier 1314 multiplies the output of the sixth adder 1312 by a first decay constant δ ⁰ .

An eighth multiplier 1316 multiplies the output of the sixth multiplier 1310 by the input signal X. A seventh adder 1318 adds the output of the eighth multiplier 1316 and subtracts the output of the seventh multiplier 1314 from the delayed first expectation by the new first expectation W. To form ⁰ .

The second expectation modifying means 1006 (1) comprises means 1302 (1) for determining the increase term, and means 1304 (1) for determining the decrease term of the second expectation. The enlargement term determining means 1302 (FIG. 1) comprise a fourth delay block 1320, which generates a second delayed expectation W ¹ _p . The enabling means 1010 (1) for the second expectation receive the first expectation W ⁰ , the delayed second expectation W ¹ _p, and the interaction signals h ^N1 (N refers to the number of signals and 1 refers to the expectation level) and generates the enable signal P ^1. The enable generating means 1010 will be explained below, the enable signal p ¹ is multiplied by a second rise constant α ¹ by a ninth multiplier 1322.

The reduction term determining means 1304 (1) for the second expectation comprise an eighth adder 1324 for subtracting the third expectation W ² _p from the second expectation W ¹ . A tenth multiplier 1326 multiplies the output of the eighth adder 1324 by a second decay constant δ ¹ .

A ninth adder 1328 adds the output of the ninth multiplier 1322 and subtracts the output of the tenth multiplier 1326 from the delayed second expectation W ¹ _p to form the new second expectation W ¹ .

The third expectation modifying means 1006 (2) comprises means 1302 (2) for determining the increase term, and means 1304 (2) for determining the decrease term of the third expectation. The enlargement term determining means 1302 (2) comprise a fifth delay block 1330 which generates a delayed third expectation W ² _p . The enable-generating means 1010 (2) for the third expectation receive the second expectation W ¹ , the delayed third expectation W ² _p and the interaction signals h ^M2 (M relates to the number of signals and 2 relates to the expectation level) and generate them the enable signal P ² . The enable signal P ² is multiplied by a third rise constant α ² by an eleventh multiplier 1332.

The reduction term determining means 1304 (2) for the third expectation comprises a tenth adder 1334 for subtracting the equilibrium expectation WE from the third expectation W ² _p . A twelfth multiplier 1336 multiplies the output of the tenth adder 1334 by a third fall constant δ ² .

An eleventh adder 1338 adds the output of the eleventh multiplier 1332 and subtracts the output of the tenth multiplier 1336 from the delayed third expectation W ² _p to form the new third expectation W ² .

With reference to Fig. 22, the enable signal generating means 1010 will be further described according to an embodiment of the present invention. The enable signal generating means 1010 comprise a multiplying bar 1402 and a hurdle signal source 1404. As shown in FIG. 23, the hurdle signal source 1404 comprises a plurality of multipliers 1502 (n). The multiplying bar receives the interaction ^signals h ^1..N and a signal (+1 or -1) from a register or table 1504. In digital representations, the multiplication by -1 can be carried out using the two's complement representation.

Returning to FIG. 22, the outputs of the multiplier 1402 bar added to each other by a twelfth adder 1406. An adder 1408 will pull the gate signal g ^l from the output of the twelfth adder 1406 from. The output of adder 1408 is connected to a first condition block 1410. A first decision block 1412 receives the previous expectation W ^l-1 . A second decision block 1414 receives the previous value of the current expectation W ^l _p . A decision block will pass the input labeled YES if the condition in the condition block is true. Therefore, if the output of the thirteenth multiplier 1408 is greater than zero, the first and second condition blocks 1412, 1414 become the previous expectation W ^l-1 and the previous value of the current expectation W ^l _p, respectively. The first and second decision blocks 1412, 1414 will pass a zero and a one, respectively, if the output of adder 1408 is not greater than zero. A thirteenth adder 1416 subtracts the output of the second decision block 1414 from the output of the first decision block 1412.

With reference to FIGS . 24-47, the expectations W ⁰ , W ¹ , W ^{2 of} an adapter with three levels and non-linear gains are shown for a number of different input signals. In each example, the first climb constant α ⁰ and δ ^{0 are set} to 0.25 and 0.125, respectively. The second rise constant and the second decay constant α ¹ and δ ¹ are 0.0625 and 0.0312, respectively. The third rise constant and the third decay constant α ² and δ ² are 0.156 and 0.0078, respectively. The maximum expectation W ^M is set to one and the equilibrium expectation W ^E is set to zero. The first, second and third expectations W ⁰ , W ¹ , W ² start at zero.

For a three-level adapter with non-linear ver Strengthening in a discrete system becomes the first expectation determined by:

W⁰ (t) = W⁰ (t-Δt) + X (t) · α⁰ · [W ^M -W⁰ (t-Δt)] - δ⁰ · [W⁰ (t-Δt) - W¹ (t-Δt)] (Eq 9A)

The second expectation is determined by:

W¹ (t) = W¹ (t-Δt) + α¹ · [W⁰-W¹ (t-Δt)] - δ¹ · [W¹ (t-Δt) - W² (t-Δt)] (Eq. 9B)

The third expectation is determined by:

W² (t) = W² (t-Δt) + α² · [W¹-W² (t-Δt)] - δ² · [W² (t-Δt) - W ^E ] (Eq. 9C)

In Fig. 24 an input signal of high intensity and short duration is applied to the Adaptrode with three levels. The first expectation W ⁰ rises while the input signal is not zero and falls when the input signal is zero. Starting with time t = 5, when the input signal does not become zero, the second expectation W ¹ also begins to rise. However, the enlargement term of the second expectation W ^{1 is} a function of the first expectation W ⁰ . Therefore, the second expectation W ¹ continues to increase after the input signal becomes zero due to the first expectation W ⁰ . In fact, while the first expectation is greater than the second expectation, the increased term of the second expectation will be positive. It should be noted at this point that the equilibrium expectation of the first expectation W ⁰ for a multi-level adapter is the following expectation, ie the second expectation W ¹ . The lower expectation of the first expectation is therefore limited by the second expectation.

Similarly, the magnification term of the third expectation is a function of the second expectation. Therefore, as shown in Fig. 24, the third expectation continues to increase while the second expectation is large enough to raise the third expectation.

In Figs. 25 and 26, an input signal is intensity of high Inten and applied medium or long duration at the Adaptrode with three levels. Both diagrams show that the second expectation W ^{1 will} increase as long as the first expectation W ^{0 is} large enough to make the enlargement term of the second expectation W ¹ larger than the decrease term.

The diagrams also show the longer-term memory capabilities of a multi-level adapter. Hence the second and third expectations from different time domains. The means that every additional level or expectation of a more level adapter has a slower response time. Although the first expectation changes relatively quickly as a result of a Input signal, it is due to the following expectations bound.

Although the first expectation W ⁰ becomes saturated due to a continuous input signal, the subsequent expectation will actually increase at a much slower rate. As shown in Fig. 27, the second and third expectations W ¹ , W ^{2 will} continue to rise, but at a much lower rate, while the first input expectation is saturated. When the input signal finally goes back to zero, the first expectation will begin to drop at a rate determined by the decrease term. However, it will be limited by the second expectation. The second expectation drops at a much slower rate so that it will keep the first expectation high depending on how long the input signal duration was non-zero and therefore how high the second expectation was driven. Similarly, the third expectation will keep the second expectation high or prevent it from decaying.

For a given expectation (W ¹ ), the subsequent expectation (W ^l-1 ) also has an impact on the expiration rate; This is illustrated by Fig. 28, in which a stochastic input signal is applied to an adapter with three levels. The input signal has the effect of driving the level of the first expectation W ⁰ up. The second expectation W ¹ is driven by the first expectation, and the third expectation is driven by the second expectation. If the input signal is zero, the reduction expectation term of the first expectation is greater than the increase term and therefore the net modification of the first expectation is negative. However, the reduction term of the first expectation is a function of the difference between the first expectation and the second expectation. As shown, the lower the decay rate for the first expectation, the higher the second expectation.

This is shown more clearly in Fig. 29, in which a sequential input signal is applied to a three-level adapter. In a first time period, the input signal becomes non-zero (1.0). During a second period, the first expectation decreases, at a rate defined by the first decay constant and the current value of the second expectation. During a third period, the input signal again becomes non-zero and the first expectation increases. In a fourth period, the input signal becomes zero again and the first expectation decreases. However, the decay rate in the fourth period is lower than the decay rate in the second period because the second expectation was driven to a higher level by the first expectation.

In FIGS. 30-35, expectations are W ^0, W ^1, W ² of a Adap trode with three levels and non-linear gains shown for a number of different input signals. Furthermore, the second expectation is gated or limited by an enable signal P.

The first expectation is determined by Equation 9A. The second expectation is determined by:

W¹ (t) = W¹ (t-Δt) + P · α¹ · [W⁰-W¹ (t-Δt)] - δ¹ · [W¹ (t-Δt) - W² (t-Δt)] (Eq.10A)

The third expectation is determined by:

W² (t) = W² (t-Δt) + α² · [W¹-W² (t-Δt)] - δ² · [W² (t-Δt) - W ^E ] (Eq. 10B)

In these examples, enable signal P is simply one Zero or one. However, the enable signal take on any value.

The input signals used in the examples in Figures 30-35 are the same as the input signals for the corresponding Figures 24-29. In each example, the enable signal is initially zero and at a predetermined time the enable signal becomes one. It should be noted that in each figure the second and third expectations do not "learn" or increase until the enable signal is one. It should also be noted that the enable signal does not affect the rate of decay of the second and third expectations.

With reference to FIGS. 36-47, the expectations W ⁰ , W ¹ , W ^{2 of} an adapter with three levels, the first expectation of which has a linear gain, are shown, for a large number of different input signals. In each example, the first rise constant α ^{0 is set} to 4 and the decay constant δ ⁰ is set to 0.125. The second rise and fall constants α ¹ , δ ¹ are set to 0.0625 and 0.0312, respectively. The third rise and fall constants α ² , δ ² are set to 0.0156 and 0.0078, respectively.

In FIGS. 36-47, the input signals 29-35 are similar to the input signals of FIG.. In Figs. 2-47, the second and third requirements by the enable signal P 30-35 are gated or restricted, as shown in Fig.. However, the enable signal is either zero or one (the enable signal is shown on a scale up to 30 for purposes of illustration). The results are similar to that shown by the non-linear models.

Referring to Fig. 48, an adaptrode based neuron or neuristor 2002 is shown. Neuristor 2002 includes an array of input adapters. The input adapter assembly 2004 includes a plurality of adaptrodes. The number of adaptrodes is generally indicated by J. Each input adaptrode 2004 (j) has an associated input signal X ^j .

The result signal r ^{j of} each input adapter 2004 (j) can be regarded as a replacement for the input signal, multiplied by the weighting signal of a standard neuron.

The neuristor 2002 comprises means 2006 for initializing and Generating a threshold signal T.

In one embodiment, they include the threshold signal generating means 2006 means for initializing the Threshold signal at a predetermined value.

In a second embodiment, the threshold signal a constant.

In a third embodiment, the smolder comprises lens signal-generating means 2006 an output adapter 2008 and a threshold processor 2010. The threshold signal is one Function of a threshold modifying signal. The Threshold-modifying signal can be the first expectation or the result signal of the output adapter. The Threshold signal generating means 2006 do this Integrate the output of the neuristor over time. The Output adapter can be replaced by the result signal (or Result signals) gated by one or more other adaptrodes or be groped.

The neuristor 2002 comprises means 2012 for receiving the input adaptrode result signal r ^j and the threshold signal T and for generating an output signal 0 in response thereto.

With reference to Fig. 49, the output signal generating means comprise means for adding 2012 of the result signal and for generating a sigma-signal. The sigma signal is compared with the threshold signal T. The output signal generating means 2012 comprise a second multiplying bar 2102 and a character register or table 2104. The output of the character register 2104 is a +1 or a -1.

Referring to Fig. 50 includes the characters source 2014. Zei chenregister or table 2202 and a plurality of Mul tiplizierern 2204 (j). The input of an adapter with an associated +1 in the character register 2204 is referred to as EXCITATORY (EXCITING) and the input of an adapter with an associated -1 in the character register 2204 is referred to as INHIBITORY (HEMMEND).

Returning to FIG. 49, the means 2012 generating output signals include an addition and output processor. A fourteenth adder 2106 adds the result signals r ^j and subtracts the threshold signal from the result signals. A third condition block 2108 compares the output of the four tenth adder 2106 with a predetermined value (0). A third decision block 2110 generates the output signal X ₀ . In a preferred embodiment, third decision block 2110 generates a one if the condition in condition block 2108 is true and a zero if the condition in condition block 2108 is false.

Returning to FIG. 48, the first expectation (or the result signal), each Eingabeadaptrode (A ^j) is used as an interaction signal (h) in the following Adaptrode (A ^{j + 1),} as shown by the arrows. It should be recognized that the keying of each adapter by the preceding adapter is only for explanatory purposes.

Generally, signals between input adapters are used for Keys can be used as local or local interactions designated signals. The result signal of the  Output adapter 2008 available for keying, as through the Arrows is shown.

Referring to Fig. 51 is a Blockdigramm of Schwellenpro zessors 2010 in accordance with an embodiment of the present invention. The threshold processor 2010 provides three types of threshold signals. The threshold signal is defined by T Type. If T Type is one, then the threshold signal is a constant with a size of T ^C. If T Type is two, then the threshold signal T is the difference between the input of the threshold processor 2010 multiplied by a constant and the constant T ^M. In one embodiment, the input to threshold processor 2010 is the first expectation of output adapter 2008 (W ⁰⁰ ). In a second exemplary embodiment, the input to the threshold processor 2010 is the result signal of the output adapter 2008. If T Type is three, then the threshold signal is a linear function of W ⁰⁰ .

The threshold processor 2010 includes fourth, fifth and sixth decision blocks 2302, 2306, 2310 and fourth, fifth and sixth control blocks 2304, 2308, 2312. A fourteenth multiplier 2314 multiplies the first expectation of the output adaptrode 2308 by a constant β. A fifteenth adder 2316 subtracts the output of the fourteenth multiplier 2314 from the constant T ^C.

Referring to FIG. 52 there is shown a graphical representation of the first and second Eingabeadaptrode Eingabeadaptrode, is used in the first Adaptrode for clocking the second He maintenance of the second Adaptrode. Each expectation is generally labeled W ^ab , where a refers to the expectation level and b to the input adapter.

Typically, neural networks can be created using a Variety of technologies to be implemented. These include "Standard" technologies such as integration on a very large scale (very large scale integration VLSI), software simulators and  Development systems, as well as special purpose hardware systems and "newer" technologies such as optical processors.

Industrial applicability. With reference to the drawings and in operation, the present invention sees a more funda mental learning ability in neural networks. The learnable capabilities of the adapter are not based on associativity per se, but rather on encrypting information contained in the time-varying signal along a single one Input channel is inherent. The encryption process is time explicit and indeed works in multiple times rods so that information for each of these time domains is recorded. Such a treatment of information ver encryption provides an alternative to return loops when implementing short and long-term memories. The adap trode is not inherently dependent on the association with another Channel to form a memory.

Associativity, however, can result from the introduction of the gate or Gate function that prevents encryption, if there is no second signal. Associations between two or more input signals or a combination of input and output signals can be achieved.

With reference to FIGS. 54-60 Adaptroden have been used to construct an associative network 2802nd The network 2802 comprises first and second neuristors 2804, 2806, each with three input adapters and one output adapter (A ¹⁰ , A ¹¹ , A ¹² , A ¹³ , A ²⁰ , A ²¹ , A ²² , A ²³ ). For reference, a legend for Figures 54-60 is shown in Figure 53. The legend shows the symbols for three types of inputs: inhibitory - 5302, learning-arousing - 5304 and non-learning-arousing - 5306.

The enable key is controlled by a local interaction between the input adapters A ¹¹ and A ¹² and between A ²³ and A ²² . The latter adaptrode, A ¹² and A ^22, is in any case the learning adapter within the relevant neuristor (as indicated by the double-plate connections). That is, the rise constants for the second and third expectations of the adaptrodes A ¹² and A ²² are not zero and the second and third expectations of the adaptrodes A ¹¹ , A ¹³ , A ²¹ and A ²³ are zero.

Thresholds have been set and set to a value sufficient to delay the output. The neuristors 2804, 2806 mutually inhibit each other that as soon as a neuristor gains an advantage it will tend to become the only responsive entity. This is achieved by A ¹¹ and A ²¹ , which have inhibitory responses, as indicated by the circular input.

The inputs UCS ¹ and UCS ² are accepted as two opposite (but not necessarily mutually exclusive) stimuli or stimuli from the external environment. UCR ¹ and UCR ² are mutually exclusive answers. The input labeled CS is available to both neuristors, as shown. Before any exposure, neither A ¹² nor A ²² encrypted any trace of memory.

For explanatory purposes, it is helpful to consider network 2802 as a simple robot controller. Here the output behavior is either to seek a stimulus (UCR ² ) or to avoid it (UCR ¹ ) if either a reward UCS ² for example for the availability of power for recharging batteries or a punishment UCR ¹ , for example the presence of high ones Temperatures associated with the corresponding stimulus.

In this case the stimulus (CS) is called the presence adopted a light source. The goal is to get the robot too train to asso light with the availability of power adorn so that he can find the light source on his own initiative becomes. It should be noted that it was just as possible would be to decide that the network is trained, light and To associate heat.  

After several attempts (the number required depends on the size of the signals, their duration and their phase relations) of common exposures, namely CS and UCS ² , this is actually achieved in 08355 00070 552 001000280000000200012000285910824400040 0002004215179 00004 08236 Fig. 55. It should be noted that the suspensions need not be at exactly the same moment (ie exactly in phase). If the CS exposure occurs just before that of UCS ² - with the time window being determined by the αs and δs in the adaptrodes - the occurrence of UCS ² will still create the enable in A ²² .

The signals offered to network 2802 are stochastic and noise (approximately 10% non-Gaussian noise). After only three attempts to present together at UCS ² and CS in the middle, the network begins to appear as Fig. 55. As can be seen, some enabling has occurred at A ²² as the result of local interaction between A ²¹ and A ²² .

Network 2802 includes a one-to-one inhibitory connection to introduce mutual inhibition. This is necessary to enforce the rule of mutual exclusion of the answer. However, it should be noted that the rule is a blurry rule in which some signal "leakage" from CS and UCS ^{1 is} possible when UCS ^{1 is} slightly active, which can occur. In general, the output of UCS ^{2 will} dampen the output of UCS ¹ and vice versa, but this is not without limitation.

The shaded areas in each adaptrode represent the levels for each expectation, as shown in the legend. It should be noted that each subsequent expectation overlaps the previous expectation. The relationships W ⁰ W ¹ W ^{2 are also} always true.

Dark arrows show that this line is "active". Therefore, the dark arrow in the second neuristor 2806 shows a strong local interaction effect of A ²³ on the learning behavior of A ²² . The arrow in the first neuristor 2804 from A ¹¹ to A ¹² represents a weak interaction and thus A ¹¹ cannot contribute enough response to allow enabling in A ¹² .

FIG. 56 illustrates the state of the network in 2802 after the network has aged in 2802 for some time without any input. As can be seen, the Adaptrode A ^{22 shows} some residual memory of the association to which it was exposed. As a result of this remainder, the network has established a relationship between the associated signals that can be recalled, as shown in FIG. 57. In Fig. 57, a signal is supplied only at CS. This corresponds to the situation that the robot only sees light. As a result, enables lichung of A ^22, the input signal at A is ^22, the first maintenance W He was ⁰ Speeding up to a higher value than in the first expectancy of A reaches ^12th In turn, this is enough to drive an output signal at UCR ² , resulting in an inhibitory signal that attenuates the first neuristor 2804.

In this way the network has learned 2802, CS, for itself itself a meaningless signal, as an announcement of a meaningful signal (in this case power available speed). The extent to which these two signals in the past has occurred together certain degree of reliability of the predictability of the Association result. The robot can learn to turn lights looking for when he has little power, and becomes a certain Put a degree of trust that this association is correct becomes.

So far the network 2802 works as you would from any asso might expect a citative learning rule. Except for something zigzig way, in time explicitly and through the learning process Gain is dynamic, the network reaches 2802 in essential what any other association based new network would reach.  

But now a more difficult consideration is made. Assume, for a short time interval, there is an opposite relationship between UCS ¹ and CS, as can happen, for example, if a fire breaks out in the vicinity of the robot. First, one would have to deal with the appropriate response; that is, the robot should avoid the fire.

Second, means are needed to determine meaning this opposite condition. Is it just noise or poses is it a more permanent change in the nature of the environment?

Third, in the event that this new relationship is over is going, the robot certainly not the previous asso forgotten because it is quite likely that the old relationship will be the norm again in the future.

In FIGS. 58-60, the progress of encrypting ge gent piece Asssoziation is shown which, since they represent different time domains can be easily stored in the network without interference. It will be seen that network 2802 responds correctly and correctly determines the level of importance that must be attached to the new conditions.

In FIG. 58, the network in 2802 a strong and some lingering charm UCS ¹ is exposed, which stands in contrast to previous experience. The system will initially begin to trigger the second neuristor 2806 as if it were responding to the availability of power. A short time is required for the first expected value W ⁰ of A ¹¹ to build up to a level which is sufficient to overdrive the output of the second neuristor 2806 via the inhibiting connection at A ²¹ . However, due to the greater actuation of the first neuristor 2804, it wins the competition, which leads to the correct answer (avoidance) in UCR ¹ . Note that A ^{22 is} the enable W ² , whereas A ^{12 is} the enable only for W ¹ .

If only the CS input is now presented to the network, it will respond for a short time, albeit weakly, with an output at UCR ¹ as in Fig. 59. The reason is that the second expectation W ¹ of A ^{12 is} on a level has risen that is just slightly larger than that of the second expectation W ¹ of A ²² . This will continue until W ¹ of A ^{12 has} dropped to the same level as W ¹ of A ²² . At this time, speaking will be ambiguous. The ambiguity will not last long. Since the exposure to the opposite condition was short and W ² of A ¹² did not increase, W ¹ of A ^{12 will} continue to fall below the level of W ¹ of A ²² . At this time, network 2802 will respond with the original UCR ² response since the older association reappears as dominant, as shown in FIG. 60.

Other aspects, goals and advantages of the present invention can from the study of drawings, revelation and the attached claims.

In summary, the invention provides the following: A synaptic processor for use in a neural network generates a result signal as a result of an input signals. The synaptic processor initializes an on expected and receives an input signal. The synaptic Processor determines a net modification of the input expectation. The net modification of the input expectation has an increase tion term and a reduction term. The enlargement term is determined as a function of the input signal. The Ver smaller term is independent of the size of the input signal and is a function of a decay constant. The synaptic Processor generates a result signal as a result of the inputs maintenance.

Claims (47)

1. A synaptic processor for generating a result signal as a result of an input signal, which comprises:
Means for initializing an input expectation, the expectation being limited by a maximum expectation and an equilibrium expectation;
Means for receiving the input signal and modifying the input expectation, the net modification of the input expectation having an increase term and a decrease term, the increase term being a function of the input signal and a first increase constant, and wherein the decrease term is a function of a first decrease constant, and wherein first rise and fall constants are independent; and
Means for generating the result signal in response to the input expectation. 2. The synaptic processor of claim 1, wherein the result signal is a linear function of input expectation. 3. The synaptic processor of claim 1, wherein the result signal a non-linear function of input expectation is. 4. Synaptic processor according to claim 1, wherein the Verklei is proportional to the difference between the Input expectation and equilibrium expectation. 5. The synaptic processor of claim 1, wherein the ver magnification term is proportional to the difference between the input expectation and the maximum expectation. 6. Synaptic processor according to claim 1, with means for Modify equilibrium expectation, where the Net modification of equilibrium expectation a second  Enlargement term and a second reduction term , the second magnification term being a function the input expectation and a second rise constant and the reduction term is a function of a second decay or decay constant. 7. The synaptic processor of claim 1, wherein the Input expectation modifying means include means to the input expectation at discrete time intervals of t to bring it up to date. 8. Synaptic processor for generating a result signal as a result of an input signal, which comprises:
Means for initializing first and second expectations, the first and second expectations being limited by a maximum expectation and an equilibrium expectation;
Means for receiving the input signal and modifying the first expectation due to the input signal and the second expectation;
Means for modifying the second expectation as a result of the first expectation; and
Means for generating the result signal as a result of the first expectation. 9. The synaptic processor of claim 8, wherein the expectations have the following relationships: W ^M W⁰ W¹ W ^E , where
W ^M = the maximum expectation
W ⁰ = the first expectation
W ¹ = the second expectation and
W ^E = equilibrium expectation
is. 10. The synaptic processor of claim 8, wherein the first Expectation is a function of the size of the input signal.   11. The synaptic processor of claim 8, wherein the first Expectation modifying means include: Reduce the first expectation as a result of the second Expectation. 12. The synaptic processor of claim 11, wherein the first expectation reducing means include means for Determine the difference between the first and second Expectations and to reduce the first expectation due to the difference and a first decay constant. 13. The synaptic processor of claim 8, wherein the first Expectation modifying means include: Increase the first expectation due to the size of the Input signals and the maximum expectation. 14. The synaptic processor of claim 13, wherein the First Expectant Enhancing Funds include Determine the difference between the first and the maximum expectation and to increase the first expectation due to the input signal, the difference and a first Rise constants. 15. The synaptic processor of claim 8, wherein the die second expectation modifying means include Increase the second expectation as a result of the first Expectation. 16. The synaptic processor of claim 13, wherein the second expectation-increasing means include means for Determine the difference between the second and first er maintenance and to increase the second expectation as a result the difference and a second rise constant. 17. The synaptic processor of claim 8, wherein the die second expectation modifying means include  Reducing the second expectation due to the Equilibrium expectation. 18. Synaptic processor according to claim 8 with means for Modify equilibrium expectation, where the Net modification of equilibrium expectation one Has enlargement term and a reduction term, where the magnification term is a function of the second Expectation and a rise or rise constant is and the reduction term being a function of a second Decay constant. 19. Synaptic processor for generating a result signal as a result of an input signal, which comprises:
Means for initializing a first expectation and at least a second expectation, the first and second expectations being limited by a maximum expectation and an equilibrium expectation, the number of second expectations being denoted by L;
Means for receiving an input signal and modifying the first expectation due to the input signal;
Means for modifying one of the second expectations, the second expectation denoted by W ¹ and being a function of: the first expectation if l = 1 or otherwise a previous second expectation W ^l-1 ; and
Means for generating the result signal as a result of the first expectation. 20. Synaptic processor according to claim 19, with means for Receipt of a second input signal and for it responsively generating an enable signal, and where the second expectation is a function of Enable signal is.   21. Synaptic processor for generating a result signal as a result of a first input signal and a second input signal, which has the following:
Means for initializing first and second expectations, the first and second expectations being limited by a maximum expectation and an equilibrium expectation;
Means for generating the first input signal and modifying the first expectation due to the first input signal and the second expectation;
Means for receiving the second input signal and responsively generating an enable signal;
Means for modifying the second expectation due to the first expectation and the enable signal, and
Means for generating the result signal as a result of the first input signal and the first expectation signal. 22. The synaptic processor of claim 21, wherein the first Expectation is a function of the size of the input signal. 23. The synaptic processor of claim 21, wherein the ini Initialization means comprise means for initializing a Gate or touch signal, and being the enable signal is generated as a result of the second input signal and Tactile signal. 24. The synaptic processor of claim 23, wherein the the Possibility signal generating means comprise means for Determine the key signal and the second input signal. 25. The synaptic processor of claim 24, wherein the second input signal determining means comprise means for receiving a variety of interaction signals and for adding the interaction signals. 26. The synaptic processor of claim 23, wherein the enable clearing signal is in a permission state if that  second input signal exceeds the key signal, and otherwise it is in a prevention state. 27. The synaptic processor of claim 26, wherein the die second expectation modifying means include Increase the second expectation as a result of the first Expectation when the enabling signal in the Is state of permission. 28. Processing element comprising:
a first synaptic processor for generating a first result signal as a result of a first input signal, which has the following:
Means for initializing an input expectation, the first input expectation being limited by a first maximum expectation and a first equilibrium expectation;
Means for receiving the first input signal and responsively modifying the input expectation, the net modification of the input expectation having a first increase term and a first decrease term, the first increase term responsive to the first input signal, and wherein the first decrease term is a function of the first equilibrium expectation; and
Means for generating the first result signal based on the input expectation;
a second synaptic processor for generating a second result signal as a result of the second input signal, comprising:
Means for initializing first and second expectations, the first and second expectations being limited by a second maximum expectation and a second equilibrium expectation;
Means for receiving the second input signal and modifying the first expectation due to the second input signal and the second expectation;
Means for receiving the first result signal and responsively generating an enable signal;
Means for modifying the second expectation due to the first expectation and the enable signal; and
Means for generating the second result signal as a result of the second input signal and the first expectation. 29. The processing element of claim 28, wherein the first synaptic processor means for modifying the first equilibrium expectation, the net modification the first equilibrium expectation a second Enlargement term and a second reduction term , the second magnification term being a function the input expectation and a second climb constant and wherein the second reduction term is a function of a second decay constant. 30. The processing element of claim 28, wherein the first Expectation modifying means include: Determine a first net modification and add it the first net modification to the first expectation, where the first net modification a third enlargement term and has a third reduction term, the third magnification term a function of the first expectation and a third climb constant, and wherein the third Reduction term a function of the first expectation and a third decay constant, and which is the second expectation modifying means include Determining a second net modification of the second Expectation and to add the second net modification to the second expectation, the second net modification a fourth magnification term and a fourth Has reduction term, the fourth Magnification term a function of the first expectation and is a fourth rise constant and the fourth  Reduction term a function of the first expectation and a second decay constant. 31. Processing element according to claim 28, with means for Modify the second equilibrium expectation, where the Net modification of the second equilibrium expectation one fifth enlargement term and a fifth Has reduction term, the fifth Magnification term a function of second expectation and is a fifth climb constant and is the fifth Reduction term a function of a fifth Decay constant. 32. Processing element comprising:
Means for generating a threshold signal;
a synaptic input processor for generating a result signal based on an input signal, the synaptic input processor comprising:
Means for initializing an input expectation, the input expectation being limited by a maximum expectation and a balance expectation;
Means for receiving the input signal and for determining the magnitude of the input signal;
Means for modifying the input expectation, the net modification of the input expectation having a first increase term and a first decrease term, the first increase term responsive to the size of the input signal, and wherein the first decrease term is independent of size and a function is an expiration constant; and
Means for generating the result signal based on the input expectation; and
Means for receiving the result signal and the threshold signal and responsively generating an output signal. 33. Processing element according to claim 32, with means for Modify equilibrium expectation, where the Net modification of equilibrium expectation a second Enlargement term and a second reduction term , the second magnification term being a function the input expectation and a second climb constant, and wherein the second reduction term is a function of a second decay constant. 34. Processing element according to claim 32, wherein the thresholds generating means comprise means for initializing a Threshold constants, for receiving a threshold modifying signal, to multiply the threshold modifying signal with a constant and to Subtract the product from the threshold constant. 35. Processing element according to claim 32, wherein the thresholds signal generating means comprise an output processor to receive the output signal and to respond to it Generate a threshold modifying signal. 36. Processing element comprising:
Means for generating a threshold signal;
at least two synaptic input processors for generating respective result signals as a result of corresponding input signals, each synaptic input processor having the following:
Means for initializing an input expectation, the input expectation being limited by a maximum expectation and a balance expectation;
Means for receiving the input signal and for determining the size of the input signal;
Means for modifying the input expectation, the net modification of the input expectation having an increase term and a decrease term, the increase term being responsive to the size of the input signal, and the decrease term being independent of size and a function of an expiration constant;
Means for generating the result signal based on the input expectation;
Means for receiving the result signals, adding the result signals and responsively generating a sigma signal; and
Means for receiving the sigma signal and the threshold signal and responsively generating an output signal. 37. Processing element according to claim 36, wherein the thresholds generating means comprise means for initializing a Threshold constants, for receiving a threshold modifying signal, to multiply the threshold modifying signal with a constant and to Subtract the product from the threshold constant. 38. Processing element according to claim 36, wherein the Output signal generating means comprise means for Comparing the threshold and sigma signals and where that Output signal has a first state in which the sigma Signal is greater than the threshold signal, and being that Otherwise, the output signal has a second state. 39. Processing element according to claim 36, wherein the thresholds signal-generating means comprise an output processor to receive the output signal and to respond to it Generate a threshold modifying signal. 40. Processing element comprising:
Means for generating a threshold signal;
a synaptic input processor for generating a result signal as a result of an input signal, the synaptic input processor comprising:
Means for initializing first and second expectations, the first and second expectations being limited by a maximum expectation and an equilibrium expectation;
Means for receiving the input signal and modifying the first expectation due to the input signal and the second expectation;
Means for modifying the second expectation as a result of the first expectation; and
Means for generating the result signal based on the second expectation; and
Means for receiving the threshold signal and the result signal and responsively generating an output signal. 41. Processing element according to claim 40, wherein the the first Expectation modifying means include: Determine a first net modification and add it the first net modification to the first expectation, where the first net modification has a first increase term and has a first reduction term, the first magnification term a function of the first expectation and a first climb constant, and wherein the first Reduction term a function of the first expectation and a first decay constant, and the means modifying the second expectation means include to determine a second net modification of the second expectation and to add the second net modes to the second expectation, the second Net modification a second enlargement term and one has second reduction term, the second Magnification term a function of second expectation and is a second rise constant and the second Reduction term a function of second expectation and a second decay constant. 42. Processing element according to claim 40, with means for Modify the equilibrium expectation, taking the net modification of the equilibrium expectation one Has enlargement term and a reduction term,  where the magnification term is a function of the second Expectation and a climb constant, and being the Reduction term a function of a second Decay constant. 43. Processing element according to claim 40 with means for Receive a second input signal and respond to it speaking generating an enable signal, the second expectation a function of the enable signal is. 44. Processing element according to claim 40, wherein the smoldering Lens signal generating means comprise an output processor to receive the output signal and to respond to it Generate a threshold modifying signal. 45. Processing element comprising:
Means for generating a threshold signal;
at least two synaptic input processors for generating respective result signals as a result of respective input signals, each of the synaptic input processors having the following:
Means for initializing first and second expectations, the first and second expectations being limited by a maximum expectation and an equilibrium expectation;
Means for receiving the input signal and modifying the first expectation due to the input signal and the second expectation;
Means for modifying the second expectation as a result of the first expectation; and
Means for generating the result signal based on the first expectation; and
Means for receiving the result signals, adding the result signals and responsively generating a sigma signal; and
Means for receiving the sigma signal and the threshold signal and responsively generating an output signal. 46. Processing element according to claim 45, wherein the first Expectation modifying means include: Determine a first net modification and add it the first net modification to the first expectation, where the first net modification has a first increase term and has a first reduction size, the first magnification term a function of the first expectation and a first climb constant, and wherein the first Reduction term a function of the first expectation and a first decay constant, and the means modifying the second expectation Comprise means for determining a second Net modification of the second expectation and for adding the second net modification to the second expectation, the second net modification being a second Enlargement term and a second reduction term , the second magnification term being a function the second expectation and a second climb constant and wherein the second reduction term is a function of the second expectation and a second decay constant. 47. Processing element according to claim 45, wherein the Threshold signal generating means an output processor include for receiving and receiving the output signal responsive generating a threshold modifier Signal.